Abstarct: Face recognition methods achieve their highest accuracy when faces are captured in the frontal mode. However, the accuracy of these methods decreases when the angle of a person's face is changed relative to the camera. The objective of this research is to identify angled faces baxsed on their overall structure and facial elements. To address the challenge of face angle variation in face images, there are two approaches. One approach involves making corrections to the face's feature vector, while the other approach involves generating a frontal face baxsed on the input angled face image. However, the latter approach incurs a significant computational cost. Since face identification is performed baxsed on feature vectors, this research proposes a method that applies modifications to the feature vectors of angled face images to increase their similarity to the feature vector of the corresponding frontal face image. In a person's face, there are two categories of information: the overall structure of the face and its elements, as well as the partial features of the face. baxsed on this, the facial elements of the angled face image are segmented using the fine-tuned DeepLabv3 network, assigning a class to each element. Then, the U-Net network equipped with an Attention module transforms the image of the angled facial elements into an image representing the frontal facial elements. Finally, features are extracted from the normalized image of facial elements using a deep convolutional network to capture general information about the features. Additionally, detailed features are extracted using the pre-trained VGGFace architecture. These detailed features are combined with the overall feature vector extracted using the facial elements. Through this combination, appropriate adjustments are made to enhance the similarity to the feature vector of the frontal face image. Mut1ny dataset is used to reset the segmentation network and FERET dataset is used to modify the feature vector. The evaluation of the proposed method was carried out on a part of the FERET dataset that contains 1600 images of 200 different people and 8 angles for each person (60, 40, 25 and 15 for both sides of the face), with an average recognition accuracy of 99.81. % is obtained and also for further evaluation, a set including 2587 images from different angles was selected from the same dataset and the recognition results were obtained on average 97.64%, which indicates the good performance of the model for more and more diverse data.